A Physics-informed Neural Network for Solving Combustion Reaction Kinetics
Recently,surrogate models for solving chemical reaction kinetics based on neural net-works have been considered to be critical for accelerating simulations of combustion reaction flows.Data-driven neural networks have long been plagued by issues such as an over-reliance on data sam-pling and difficulty in ensuring robustness and generalization.To address these issues,the training of neural networks is constrained by physical principles in the present study,including the law of mass action,and the conservation of mass,energy,and elements.Compared with data-driven neural network surrogate models,the physics-informed neural network accelerates computations by 2.0～4.7 times while suppressing prediction errors in the numerical simulations of 0-D autoignitions and a 2-D Bunsen flame.Finally,based on theoretical estimates of the source term error in combustion reaction flow simulations,this study introduces several suggestions to reduce prediction errors and improve model robustness.
combustion reaction kineticsphysics-informed neural networksource term error
NETL
NSTL
万方数据
